----------------------------------------
What I'm doing at work
October 11th, 2026
----------------------------------------
I don't write about work very much. That's probably for the best.
Work is something I do out of need to provide. I have so many
passion hobbies and interest they should be the focus of my
writing here. Still, I think I'm pretty good at my job and I take
some satisfaction in doing things I think are interesting when
I can.
As a brief reminder, I work in healthcare marketing. I am
a managing partner of an agency devoted to that, and my domain is
what we call Solutions. That includes things like tech, analytics,
strategy, and media. I'm responsible for overseeing those things
at a very strategic level, setting policy, direction, staffing our
senior team, and playing a role in some of our higher-level client
relationships. Occasionally I get to play in the weeds, helping to
craft a program to address as strategic need, or even building
some custom software.
Right now I'm working on something that falls in the later
category. It's related to AI (boo). In marketing, you see, we're
very concerned with how to reach audiences with our messaging.
Whether I'm talking about a patient and teaching them about
a savings card program for a drug, or sharing disease state
information with a doctor on emerging publications, we need to be
able to get that message to them at the right moment. There's
a lot of systems involved in that, and there's no one single
pathway. One of the major emerging methods is with AI.
When people perform a search on Google, Gemini is there (with
a weak model) to give an AI answer overview. Believe it or not,
that has changed the relationship with search engines quite
considerably on its own. If you consider how many people don't go
to a traditional search engine at all anymore, preferring to ask
ChatGPT or Claude their questions, that new reality shifts even
more.
Now this isn't an all-or-nothing change. Overall, total search
volume has increased on the web over the last few years, but
click-through engagement on branded websites has gone down. The
difference is hypothesized to live in the AI middle-ground. When
an LLM gives you an answer, it doesn't give you a list of links
which it's determined to have high authority on the subject you're
searching. It gives you an amalgomation of the data it's trained
upon as a direct statement. Instead of links, it gives you
a narrative.
In the healthcare space this is dangerous for a lot of obvious
reasons. You don't want your machine telling you to eat glue on
pizza, right? So there was some public outcry when these things
started, and the AI companys responded by being responsible and
putting real safeguards in place which source only from the most
trustworthy information on the--yeah, that's bullshit. They
responded by adding a post-response guard system:
1) Identify that the question relates to a health topic
2) Inject an invisible guard prompt into the LLM to avoid giving
dangerous advice that's been identified in a previously
reported pattern.
3) Read the response text and run a second query out to find
citations from a list of trustworthy sources which seem to
related to the content being shown.
4) Profit.
Yeah, that's right. They make up the citations after the response
is written. I've got some research that shows between 50-90% of
medical queries aren't fully supported by their citations.
A Stanford study called it a "facade of trustworthiness".
So what is impacting the answers? Overwhelmingly it's social media
content, just from sheer volume of information and the structured
linking of that content being easy to use for training. It's also
frustratingly easy to use social media to poison that AI to say
what you want[0].
(HTM) [0] 404 Media article on AI poisoning
So, that sorta sucks, right? For humans, I mean. I hope it gets
better. I hope the frontier model business structures collapse
under their own weight, and data center projects fail, and we get
new leaders who think about more than their own pocket. But that's
not the focus of this post. This is about my capitalism job, where
I need to figure out how to deal with this changing reality for
marketing.
So what am I doing? I'm focusing on the answers the LLMs give us.
Not the citations they claim, not the number of times key words
appear, or the sentiment of the posts, but the actual lexicon. I'm
comparing that with our other work in Social Listening and media
reporting, SEO/SEM data, and more. I'm focused on the public
narrative being shared by humans and machines about a brand or
topic. And then I'm doing real textual analysis work on that
content to determine if it's stable across queries, wording, LLM
models, channels, and time. I'm looking at whether it's aligning
with the brand narrative the client is trying to share. Where are
the gaps?
This lexicon analysis work is fun. There's a lot of math and
figuring out what's actually useful is an interesting hard
challenge. In the end this will surface a new type of landscape
analysis for our analytics and strategy teams to dig into and plot
out new marketing efforts. How do we steer those conversations?
How do we address misinformation? Oh, and this is a real and valid
use case! My industry is often the subject of deserved criticism,
but there are legitimately people involved who want to make sure
factual, well-researched health information is reaching the
audience.
Anyway, that's my latest project. It won't be published publicly.
We'll have this as a custom tool we offer to our clients,
different than the GEO auditing tools out there, and a unique
value we can bring. Eventually the market will change again, or
others will start doing something similar, but we'll just think
about what to build next to stay ahead.